In humans the concepts of reasonable and safe actions can be taught directly via natural language. In this project we will develop an understanding of explicit information from dialogues in natural language to infer and learn safe and unsafe positions and actions. Under real-world conditions and based on a probabilistic grammar and vocabulary a robot should be able to extract labels for entities and identifiers for safe actions based on a few training samples or one shot learning.
Effective and applicable methods for language understanding to enable robots to extract information on safe behavior in different contexts and environments. The results will complement findings and methods from Project 8 by adding threat detection on word and sentence level.